Decentralized Federated Learning with Differential Privacy via Graph Consensus

Abstract

This paper presents a novel approach to decentralized federated learning that leverages graph consensus algorithms to achieve robust and privacy-preserving model aggregation. The core idea is to represent participating devices as nodes within a graph and utilize graph consensus protocols for model update exchange and aggregation. Differential privacy is integrated into this framework to protect the sensitive data residing on individual devices. We demonstrate that this combined approach offers a scalable and efficient solution for distributed learning, particularly in scenarios with a highly decentralized network topology. The proposed system addresses key challenges associated with traditional federated learning, such as the vulnerability to malicious actors and the potential for information leakage, while maintaining model accuracy. The theoretical framework outlines the convergence properties of the graph consensus algorithm and provides a rigorous analysis of the privacy guarantees afforded by differential privacy. This work contributes to the growing field of privacy-preserving distributed learning, offering a practical and adaptable solution for diverse applications. ---